International Scholarly Research Network ISRN Probability and Statistics Volume 2012, Article ID 789273, 20 pages doi:10.5402/2012/789273 Research Article Matrix Variate Pareto Distribution of the Second Kind Daya K. Nagar, 1 Lata Joshi, 2 and Arjun K. Gupta 3 1 Instituto de Matem´ aticas, Universidad de Antioquia, Calle 67, No. 53–108, Medell´ ın, Colombia 2 Department of Statistics, University of Rajasthan, Jaipur 302004, India 3 Department of Mathematics and Statistics, Bowling Green State University, Bowling Green, OH 43403-0221, USA Correspondence should be addressed to Daya K. Nagar, [email protected]Received 15 August 2012; Accepted 6 September 2012 Academic Editors: J. Jiang and C. Proppe Copyright q 2012 Daya K. Nagar et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. We generalize the univariate Pareto distribution of the second kind to the matrix case and give its derivation using matrix variate gamma distribution. We study several properties such as cumulative distribution function, marginal distribution of submatrix, triangular factorization, moment generating function, and expected values of the Pareto matrix. Some of these results are expressed in terms of special functions of matrix arguments, zonal, and invariant polynomials. 1. Introduction The Lomax distribution, also called the Pareto distribution of the second kind is given by the p.d.f. β λ 1 v λ −β1, v> 0, 1.1where shape parameter β> 0 and location parameter λ> 0. The Lomax distribution, named after Lomax, is a heavy-tail probability distribution often used in business, economics, and actuarial modeling. The standard Pareto Distribution of the second kind has λ 1 with the p.d.f. β1 v−β1, v> 0,β> 0. 1.2
21
Embed
Matrix Variate Pareto Distribution of the Second Kinddownloads.hindawi.com/archive/2012/789273.pdf · 2019. 7. 31. · The matrix variate Pareto distribution can be derived by using
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
International Scholarly Research NetworkISRN Probability and StatisticsVolume 2012, Article ID 789273, 20 pagesdoi:10.5402/2012/789273
Research ArticleMatrix Variate Pareto Distribution of theSecond Kind
Daya K. Nagar,1 Lata Joshi,2 and Arjun K. Gupta3
1 Instituto de Matematicas, Universidad de Antioquia, Calle 67, No. 53–108, Medellın, Colombia2 Department of Statistics, University of Rajasthan, Jaipur 302004, India3 Department of Mathematics and Statistics, Bowling Green State University, Bowling Green,OH 43403-0221, USA
Correspondence should be addressed to Daya K. Nagar, [email protected]
Received 15 August 2012; Accepted 6 September 2012
Academic Editors: J. Jiang and C. Proppe
Copyright q 2012 Daya K. Nagar et al. This is an open access article distributed under the CreativeCommons Attribution License, which permits unrestricted use, distribution, and reproduction inany medium, provided the original work is properly cited.
We generalize the univariate Pareto distribution of the second kind to the matrix case and giveits derivation using matrix variate gamma distribution. We study several properties such ascumulative distribution function, marginal distribution of submatrix, triangular factorization,moment generating function, and expected values of the Pareto matrix. Some of these results areexpressed in terms of special functions of matrix arguments, zonal, and invariant polynomials.
1. Introduction
The Lomax distribution, also called the Pareto distribution of the second kind is given by thep.d.f.
β
λ
(1 +
v
λ
)−(β+1), v > 0, (1.1)
where shape parameter β > 0 and location parameter λ > 0. The Lomax distribution, namedafter Lomax, is a heavy-tail probability distribution often used in business, economics, andactuarial modeling. The standard Pareto Distribution of the second kind has λ = 1 with thep.d.f.
β(1 + v)−(β+1), v > 0, β > 0. (1.2)
2 ISRN Probability and Statistics
Although a wealth of results on Pareto distribution is available in the literature (see Johnsonet al. [1]) nothing appears to have been done to define and study matrix variate Paretodistribution.
Therefore, in this paper, we define matrix variate Pareto distribution and study severalof its properties.
We will use the following standard notations (cf. Gupta and Nagar [2]). Let A = (aij)be an m × m matrix. Then, AT denotes the transpose of A; tr(A) = a11 + · · · + amm; etr(A) =exp(tr(A)); det(A) = determinant of A; ‖A‖ = norm of A; A > 0 means that A is symmetricpositive definite and A1/2 denotes the unique symmetric positive definite square root of A >0. The submatrices A(α) and A(α), 1 ≤ α ≤ m, of the matrix A are defined as A(α) = (aij), 1 ≤i, j ≤ α, and A(α) = (aij), α ≤ i, j ≤ m, respectively.
The multivariate gamma function which is frequently used in multivariate statisticalanalysis is defined by
Γm(a) =∫
X>0etr(−X)det (X)a−(m+1)/2dX
= πm(m−1)/4m∏i=1
Γ(a − i − 1
2
), Re(a) >
m − 12
.
(1.3)
The multivariate generalization of the beta function is given by
Bm(a, b) =∫ Im
0det (X)a−(m+1)/2 det (Im −X)b−(m+1)/2 dX
=Γm(a)Γm(b)Γm(a + b)
= Bm(b, a),
(1.4)
where Re(a) > (m − 1)/2 and Re(b) > (m − 1)/2. Further, by using the matrix transformationX = (Im + Y )−1Y in (1.4) with the Jacobian J(X → Y ) = det(Im + Y )−(m+1) one can easilyestablish the identity
Bm(a, b) =∫
Y>0det (Y )a−(m+1)/2 det (Im + Y )−(a+b) dY. (1.5)
The beta type 1 and beta type 2 families of distributions are defined by the densityfunctions (Johnson et al. [1]):
{B(α, β
)}−1uα−1(1 − u)β−1, 0 < u < 1, (1.6)
{B(α, β
)}−1vα−1(1 + v)−(α+β), v > 0, (1.7)
respectively, where α > 0, β > 0, and
B(α, β
)=
Γ(α)Γ(β)
Γ(α + β
) . (1.8)
ISRN Probability and Statistics 3
Recently, Cardeno et al. [3] have defined and studied the family of beta type 3 distributions.A random variablew is said to follow a beta type 3 distribution if its density function is givenby
2α{B(α, β
)}−1wα−1(1 −w)β−1(1 +w)−(α+β), 0 < w < 1. (1.9)
If a random variable u has the p.d.f. (1.6), then we will write u ∼ B1(α, β), and if thep.d.f. of the random variable v is given by (1.7), then v ∼ B2(α, β). The distribution given bythe density (1.9) will be designated by w ∼ B3(α, β). The matrix variate generalizations of(1.6), (1.7), and (1.9) are defined as follows (Gupta and Nagar [2, 4, 5]).
Definition 1.1. Anm×m random symmetric positive definite matrixU is said to have a matrixvariate beta type 1 distribution with parameters (α, β), denoted asU ∼ B1(m,α, β), if its p.d.f.is given by
det (U)α−(m+1)/2 det (Im −U)β−(m+1)/2
Bm
(α, β
) , 0 < U < Im, (1.10)
where α > (m − 1)/2 and β > (m − 1)/2.
Definition 1.2. Anm×m random symmetric positive definite matrix V is said to have a matrixvariate beta type 2 distribution with parameters (α, β), denoted as V ∼ B2(m,α, β), if its p.d.f.is given by
det (V )α−(m+1)/2 det (Im + V )−(α+β)
Bm
(α, β
) , V > 0, (1.11)
where α > (m − 1)/2 and β > (m − 1)/2.
Definition 1.3. Anm×m random symmetric positive definite matrixW is said to have amatrixvariate beta type 3 distribution with parameters (α, β), denoted asW ∼ B3(m,α, β), if its p.d.f.is given by
2mα det (W)α−(m+1)/2 det (Im −W)β−(m+1)/2
Bm
(α, β
)det (Im +W)α+β
, 0 < W < Im, (1.12)
where α > (m − 1)/2 and β > (m − 1)/2.
2. The Density Function
First we define the matrix variate Pareto distribution of the second kind.
4 ISRN Probability and Statistics
Definition 2.1. Anm×m random symmetric positive definite matrix V is said to have a matrixvariate Pareto distribution of the second kind, denoted as V ∼ Pm(β), β > (m−1)/2, if its p.d.f.is given by
Γm[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det (Im + V )−β−(m+1)/2, V > 0. (2.1)
Definition 2.2. Anm×m random symmetric positive definite matrixU is said to have a matrixvariate Lomax distribution with parameters Λ and β, denoted as U ∼ Lm(Λ, β), if its p.d.f. isgiven by
Γm[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det (Λ)−(m+1)/2 det(Im + Λ−1U
)−β−(m+1)/2, U > 0, (2.2)
where Λ is an m ×m symmetric positive definite matrix and β > (m − 1)/2.
From Definitions 2.1 and 2.2 it is clear that if V ∼ Pm(β), then for an m ×m symmetricpositive definite constant matrix Λ, Λ1/2VΛ1/2 ∼ Lm(Λ, β) and if U ∼ Lm(Λ, β), thenΛ−1/2UΛ−1/2 ∼ Pm(β).
Form = 1, thematrix variate Pareto distribution andmatrix variate Lomax distributionreduce to their respective univariate forms.
The matrix variate Pareto distribution can be derived by using independent gammamatrices. A random matrix Y is said to have a matrix variate gamma distribution withparameters Ψ (> 0) and κ (> (m − 1)/2), denoted by Y ∼ Ga(m,κ,Ψ), if its p.d.f. is givenby
etr(−Ψ−1Y
)det (Y )κ−(m+1)/2
Γm(κ)det (Ψ)κ, Y > 0. (2.3)
Theorem 2.3. Let Y1 and Y2 be independent, Y1 ∼ Ga(m, (m + 1)/2, Im) and Y2 ∼ Ga(m, β, Im).Then, Y−1/2
2 Y1Y−1/22 ∼ Pm(β).
Proof. The joint density function of Y1 and Y2 is given by
etr[−(Y1 + Y2)]det (Y2)β−(m+1)/2
Γm[(m + 1)/2]Γm(β) , Y1 > 0, Y2 > 0. (2.4)
Transforming W = Y−1/22 Y1Y
−1/22 with the Jacobian J(Y1 → W) = det(Y2)
(m+1)/2 in the jointdensity of Y1 and Y2, we obtain the joint density of W and Y2 as
etr[−(Im +W)Y2]det (Y2)β+(m+1)/2−(m+1)/2
Γm[(m + 1)/2]Γm(β) , 0 < W < Im, Y2 > 0. (2.5)
Now, the desired result is obtained by integrating Y2 using (1.3).
ISRN Probability and Statistics 5
The cumulative distribution function of V is obtained as
P(V < Ω) =Γm
[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
∫Ω
0det (Im + V )−β−(m+1)/2 dV
=Γm
[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det (Ω)(m+1)/2∫ Im
0det (Im + ΩW)−β−(m+1)/2 dW,
(2.6)
where the last line has been obtained by substituting W = Ω−1/2VΩ−1/2 with the JacobianJ(V → W) = det(Ω)(m+1)/2. Now, writing
Finally, using the integral representation of the Gauss hypergeometric function (Herz [6],Constantine [7], James [8], and Gupta and Nagar [2]), namely,
2F1(a, b; c;X) =Γm(c)
Γm(a)Γm(c − a)
∫ Im
0det(R)a−(m+1)/2
× det(Im − R)c−a−(m+1)/2det(Im −XR)−bdR,
(2.9)
where Re(a) > (m − 1)/2, Re(c − a) > (m − 1)/2, and X < Im, we obtain
P(V < Ω) =Γm
[β + (m + 1)/2
]Γm[(m + 1)/2]
Γm(β)Γm(m + 1)
det (Ω)(m+1)/2
det (Im + Ω)β+(m+1)/2
× 2F1
(m + 12
, β +m + 12
;m + 1; (Im + Ω)−1Ω).
(2.10)
The moment generating function of V is derived as
MX(Z) =Γm
[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
∫
V>0etr(ZV )det (Im + V )−β−(m+1)/2dV, (2.11)
6 ISRN Probability and Statistics
where Z (m ×m) = ((1 + δij)zij/2). Now, evaluating the above integral, we obtain
MX(Z) =Γm
[β + (m + 1)/2
]
Γm(β) Ψ
(m + 12
,−β +m + 12
;−Z), (2.12)
where the confluent hypergeometric functionΨ, withm×m symmetric matrixX as argument,is defined by the integral
Ψ(a, c;X) =1
Γm(a)
∫
R>0etr(−RX)det (R)a−(m+1)/2
× det (Im + R)c−a−(m+1)/2dR
(2.13)
valid for Re(X) > 0 and Re(a) > (m − 1)/2.
3. Properties
In this section, we give several properties of the matrix variate Pareto distribution of thesecond kind defined in the previous section.
Theorem 3.1. Let V ∼ Pm(β) and letA be anm×m constant nonsingular matrix. Then, the densityof X = AVAT is
Γm[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det(AAT
)βdet
(AAT +X
)−β−(m+1)/2, X > 0. (3.1)
Theorem 3.2. Let V ∼ Pm(β) and let H be an m ×m orthogonal matrix, whose elements are eitherconstants or random variables distributed independent of V . Then, the distribution of V is invariantunder the transformation V → HVHT . Further, if H is a random matrix, then V and H aredistributed independently.
Theorem 3.3. If V ∼ Pm(β), then, V −1 ∼ B2(m, β, (m+1)/2), (2Im+V )−1V ∼ B3(m, (m+1)/2, β)and (Im+2V )−1 ∼ B3(m, β, (m+1)/2). Further, ifW ∼ B3(m, (m+1)/2, β), then 2(Im−W)−1W ∼Pm(β).
Theorem 3.4. Let V =(
V11 V12V21 V22
), where V11 is a q × q matrix. Define V11·2 = V11 − V12V
−122 V21 and
V22·1 = V22 − V21V−111 V12. If V ∼ Pm(β), then (i) V11 and V22·1 are independent, V11 ∼ B2(q, (m +
1)/2, β − (m − q)/2) and V22·1 ∼ Pm−q(β); (ii) V22 and V11·2 are independent, V22 ∼ B2(m − q, (m +1)/2, β − q/2) and V11·2 ∼ Pq(β).
Proof. From the partition of V , we have
det(Im + V ) = det(Iq + V11
)det
(Im−q + V22·1 + V21V
−111
(Iq + V11
)−1V12
). (3.2)
ISRN Probability and Statistics 7
Now, making the transformation V11 = V11, X = V21V−1/211 and V22·1 = V22 − V21V
−111 V12 =
V22 −XXT with the Jacobian J(V11, V22, V21 → V11, V22·1, X) = det(V11)(m−q)/2 in the density of
V , we get the joint density of V11, V22·1, and X as
Γm[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det (V11)(m−q)/2 det (Im + V11)−β−(m+1)/2
× det(Im−q + V22·1 +X(Im + V11)−1XT
)−β−(m+1)/2.
(3.3)
Further, transforming Y = (Im−q + V22·1)−1/2X(Iq + V11)
−1/2 with the Jacobian J(X → Y ) =det(Im−q + V22·1)
q/2 det(Iq + V11)(m−q)/2, the joint density of V11, V22·1, and Y is derived as
Γm[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det (V11)(m−q)/2 det (Im + V11)−β−(q+1)/2
× det(Im−q + V22·1
)−β−(m−q+1)/2 det(Im−q + YYT
)−β−(m+1)/2,
(3.4)
where V11 > 0, V22·1 > 0 and Y ∈ R(m−q)×q. From the above factorization, it is clear that V11 and
V22·1 are independent, V11 ∼ B2(q, (m + 1)/2, β − (m − q)/2) and V22·1 ∼ Pm−q(β). The secondpart is similar.
Theorem 3.5. Let A be a q ×m constant matrix of rank q (≤ m). If V ∼ Pm(β), then
[(AAT
)−1/2AV −1AT
(AAT
)−1/2]−1 ∼ Pq
(β),
(AAT
)−1/2AVAT
(AAT
)−1/2 ∼ B2(q,
m + 12
, β − m − q
2
).
(3.5)
Proof. Write A = M(Iq, 0)G, where M is a q × q nonsingular matrix and G is an m × morthogonal matrix. Now,
(AV −1AT
)−1=(M
(Iq 0
)GV −1GT(Iq 0
)TMT
)−1
=(MT
)−1[(Iq 0
)Y−1
(Iq0
)]−1M−1
=(MT
)−1(Y 11
)−1M−1,
(3.6)
8 ISRN Probability and Statistics
where Y =(
Y11 Y12Y21 Y22
)= GVGT ∼ Pm(β), Y11 is a q × q matrix, and Y 11 = (Y11 − Y12Y
−122 Y21)
−1 =
Y−111·2. From Theorem 3.4, Y11·2 ∼ Pq(β), and Theorem 3.1, Z = (MT )−1Y11·2M−1 has the p.d.f.
proportional to
Γm[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det(MMT
)−βdet
((MMT
)−1+ Z
)−β−(m+1)/2
, Z > 0. (3.7)
Now, noting thatMMT = AAT andmaking the transformation S = (AAT )1/2Z(AAT )1/2 withthe Jacobian J(Z → S) = det(AAT )−(m+1)/2 in the above density, we get the desired result.The proof of the second part is similar.
From Theorem 3.5, it is clear that if V ∼ Pm(β) and a ∈ Rm, a/= 0, then aTa(aTV −1a)−1 ∼
P(β). Further, if y(m × 1) is a random vector, independent of W , and P(y/= 0) = 1, then itfollows that yTy(yTV −1y)−1 ∼ P(β) and (yTy)−1(yTVy) ∼ B2((m + 1)/2, β − (m − 1)/2).
From the above results, it is straightforward to show that if c(m × 1) is a nonzeroconstant vector or a random vector independent of V with P(c/= 0) = 1, then
cTV −1ccT(V −1 + Im
)c∼ B1
(β, 1
),
cTccT(V −1 + Im
)c∼ B1
(1, β
),
cTV −1ccTc
∼ B2(β, 1
),
cTV ccTc
∼ B2(m + 12
, β − m − 12
).
(3.8)
The expectation of V , E(V ), can easily be obtained from the above result. For any fixedc ∈ R
m, c/= 0,
E
[cTV ccTc
]= E(v), (3.9)
where v ∼ B2((m + 1)/2, β − (m − 1)/2). Hence, for all c ∈ Rm,
cTE(V )c = cTcE(v) =m + 1
2β −m − 1cTc, β >
m + 12
(3.10)
which implies that
E(V ) =m + 1
2β −m − 1Im, β >
m + 12
. (3.11)
ISRN Probability and Statistics 9
Theorem 3.6. If V ∼ Pm(β) and V = WWT , where W = (wij) is a lower triangular matrix withpositive diagonal elements, thenw11, . . . , wmm are all independent,w2
ii ∼ B2((m+ 2− i)/2, β − (m−i)/2), i = 1, . . . , m.
Proof. Making the transformation V = WWT with the Jacobian J(V → W) = 2m∏m
i=1wm+1−iii
in (2.1), the density of W is derived as
2mΓm
[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det(Im +WWT
)−β−(m+1)/2 m∏i=1
wm+1−iii , (3.12)
where wii > 0, i = 1, . . . , m and −∞ < wij < ∞ for 1 ≤ j < i ≤ m. Now, partition W as
W =(w11 0w W22
), (3.13)
wherew is an (m− 1)× 1 vector andW22 is an (m− 1)× (m− 1) lower triangular matrix. Then
det(Im +WWT
)= det
(1 +w2
11 w11wT
w11w Im−1 +wwT +W22WT22
)
=(1 +w2
11
)det
(Im−1 +W22W
T22
)
×[1 +
11 +w2
11
wT(Im−1 +W22W
T22
)−1w
].
(3.14)
Now, make the transformation
y =1
(1 +w2
11
)1/2(Im−1 +W22W
T22
)−1/2w (3.15)
with the Jacobian J(w → y) = (1 +w211)
(m−1)/2 det(Im−1 +W22WT22)
1/2 in (3.12) to get the jointdensity of w11, W22, and y as
2mΓm
[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
wm11
(1 +w2
11
)−β−1
× det(Im−1 +W22W
T22
)−β−m/2 m∏i=2
wm+1−iii
(1 + yTy
)−β−(m+1)/2.
(3.16)
From the above factorization, it is clear that w11, W22, and y are all independent, w211 ∼
B2((m + 1)/2, β − (m − 1)/2) and the density ofW22 is proportional to
det(Im−1 +W22W
T22
)−β−m/2 m∏i=2
wm+1−iii (3.17)
10 ISRN Probability and Statistics
which has the same form as the density (3.12) with m replaced by m − 1. Repeating theargument given above on the density function of W22, we observe that w2
22 ∼ B2(m/2, β −(m − 2)/2) and is independent of w33, . . . , wmm. Continuing further with the same argument,we get the desired result.
Corollary 3.7. If V ∼ Pm(β), then the distribution of det(V ) is the same as the distribution of theproduct of m independent beta type 2 variables, that is, det(V ) ∼ ∏m
i=1vi where vi ∼ B2((m + 2 −i)/2, β − (m − i)/2), i = 1, . . . , m.
Corollary 3.8. If V ∼ Pm(β), then
det(V (1))
det(V (0)
) , det(V (2))
det(V (1)
) , . . . , det(V (m))
det(V (m−1))
(det
(V (0)
)≡ 1
)(3.18)
are independently distributed. Further, for i = 1, . . . , m, det(V (i))/det(V (i−1)) ∼ B2((m+2−i)/2, β−(m − i)/2).
Theorem 3.9. If V ∼ Pm(β) and V = WWT , where W = (wij) is an upper triangular matrixwith positive diagonal elements, then w11, . . . , wmm are all independent, w2
ii ∼ B2((i + 1)/2, β − (i −1)/2), i = 1, . . . , m.
Proof. Making the transformation V = WWT with the Jacobian J(V → W) = 2m∏m
i=1wiii in
(2.1), the density of W is derived as
2mΓm
[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
det(Im +WWT
)−β−(m+1)/2 m∏i=1
wiii, (3.19)
where wii > 0, i = 1, . . . , m and −∞ < wij < ∞ for 1 ≤ i < j ≤ m. Now, partition W as
W =(W11 w0 wmm
), (3.20)
wherew is an (m− 1)× 1 vector andW11 is an (m− 1)× (m− 1) upper triangular matrix. Then
det(Im +WWT
)= det
(Im−1 +wwT +W11W
T11 wmmw
wmmwT 1 +w2mm
)
=(1 +w2
mm
)det
(Im−1 +W11W
T11
)
×[1 +
11 +w2
mm
wT(Im−1 +W11W
T11
)−1w].
(3.21)
Now, make the transformation
y =1
(1 +w2
mm
)1/2(Im−1 +W11W
T11
)−1/2w (3.22)
ISRN Probability and Statistics 11
with the Jacobian J(w → y) = (1+w2mm)
(m−1)/2 det(Im−1 +W11WT11)
1/2 in (3.19) to get the jointdensity of W11, y, and wmm as
2mΓm
[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
wmmm
(1 +w2
mm
)−β−1
× det(Im−1 +W11W
T11
)−β−m/2m−1∏i=1
wiii
(1 + yTy
)−β−(m+1)/2.
(3.23)
From the above factorization, it is clear that W11, wmm, and y are all independent, w2mm ∼
B2((m + 1)/2, β − (m − 1)/2) and the density ofW11 is proportional to
det(Im−1 +W11W
T11
)−β−m/2m−1∏i=1
wiii (3.24)
which has the same form as the density (3.19) with m replaced by m − 1. Repeatingthe argument given above on the density function of W11, we observe that w2
m−1,m−1 ∼B2(m/2, β − (m − 2)/2) and is independent of wm−2,m−2, . . . , w11. Continuing further with thesame argument, we get the desired result.
Corollary 3.10. If V ∼ Pm(β), then the distribution of det(V ) is the same as the distribution of theproduct ofm independent beta type 2 variables, that is, det(V ) ∼ ∏m
i=1vi where vi ∼ B2((i+1)/2, β−(i − 1)/2), i = 1, . . . , m.
Corollary 3.11. If V ∼ Pm(β), then
det(V(1)
)
det(V(2)
) , det(V(2)
)
det(V(3)
) , . . . , det(V(m)
)
det(V(m+1)
) (det
(V(m+1)
) ≡ 1)
(3.25)
are independently distributed. Further, for i = 1, . . . , m, det(V(i))/det(V(i+1)) ∼ B2((i + 1)/2, β −(i − 1)/2).
We conclude this section by deriving moments of det(V ) and det(Im + V )−1.
Theorem 3.12. Let V ∼ Pm(β), then
E
[det (V )r
det (Im + V )s
]=
Γm[β + (m + 1)/2
]Γm[(m + 1)/2 + r]Γm
(β + s − r
)
Γm[β + (m + 1)/2 + s
]Γm[(m + 1)/2]Γm
(β) , (3.26)
where Re(r) > −1 and Re(β + s) > (m − 1)/2.
12 ISRN Probability and Statistics
Proof. By definition
E
[det (V )r
det (Im + V )s
]=
Γm[β + (m + 1)/2
]
Γm(β)Γm[(m + 1)/2]
∫
V>0
det (V )(m+1)/2+r−(m+1)/2 dV
det (Im + V )β+(m+1)/2+s. (3.27)
Now, evaluating the above integral using (1.5), we get the result.
Corollary 3.13. If V ∼ Pm(β), then
E[det (V )r
]=
Γm[(m + 1)/2 + r]Γm(β − r
)
Γm[(m + 1)/2]Γm(β) , Re
(β − r
)>
m − 12
,
E[det (Im + V )−s
]=
Γm[β + (m + 1)/2
]Γm
(β + s
)
Γm[β + (m + 1)/2 + s
]Γm
(β) , Re
(β + s
)>
m − 12
.
(3.28)
By writing multivariate gamma functions in terms of ordinary gamma functions, expressionsE[det(V )r] and E[det(Im + V )−s] can be simplified as
E[det (V )r
]=
m∏j=1
Γ[(m + 1)/2 + r − (
j − 1)/2
]Γ(β − r − (
j − 1)/2
)
Γ[(m + 1)/2 − (
j − 1)/2
]Γ(β − (
j − 1)/2
) , (3.29)
E[det (Im + V )−s
]=
m∏j=1
Γ[β + (m + 1)/2 − (
j − 1)/2
]Γ[β + s − (
j − 1)/2
]
Γ[β + (m + 1)/2 + s − (
j − 1)/2
]Γ[β − (
j − 1)/2
] . (3.30)
Substituting r, s = 1, 2, the first and second order moments of det(V ) and det(Im + V )−1 arecalculated as
E[det(V )] =(2)m(
2β −m − 1)m
, β >m + 12
,
E[det (V )2
]=
(2)m(4)m(2β −m − 1
)m
(2β −m − 3
)m
, β >m + 32
,
E[det (Im + V )−1
]=
(2β −m + 1
)m(
2β + 2)m
, β >m − 12
,
E[det (Im + V )−2
]=
(2β −m + 1
)m
(2β −m + 3
)m(
2β + 2)m
(2β + 4
)m
, β >m − 12
,
(3.31)
where the Pochhammer notation (a)k is defined by (a)k = a(a + 1) · · · (a + k − 1), k =1, 2, . . . with (a)0 = 1.
ISRN Probability and Statistics 13
4. Results Involving Zonal and Invariant Polynomials
Let Cκ(X) be the zonal polynomial of an m × m symmetric matrix X corresponding to thepartition κ = (k1, . . . , km), k1 + · · · + km = k, k1 ≥ · · · ≥ km ≥ 0. Then, for small values of k,explicit formulas for Cκ(X) are available as (James [8])
C(1)(X) = tr(X),
C(2)(X) =13
[(trX)2 + 2 tr
(X2
)],
C(12)(X) =23
[(trX)2 − tr
(X2
)],
C(3)(X) =115
[(trX)3 + 6(trX)
(trX2
)+ 8 tr
(X3
)],
C(2,1)(X) =35
[(trX)3 + (trX)
(trX2
)− 2 tr
(X3
)],
C(13)(X) =13
[(trX)3 − 3(trX)
(trX2
)+ 2 tr
(X3
)].
(4.1)
From the above results, it is straightforward to show that
tr(X2
)= C(2)(X) − 1
2C(12)(X),
tr(X3
)= C(3)(X) − 1
4C(2,1)(X) +
14C(13)(X),
tr(X) tr(X2
)= C(3)(X) +
16C(2,1)(X) − 1
2C(13)(X).
(4.2)
For an ordered partition ρ of r, ρ = (r1, . . . , rm), r1 ≥ · · · ≥ rm ≥ 0, r1 + · · · + rm = r,Γm(a, ρ) and Γm(a,−ρ) are defined by
Γm(a, ρ
)= (a)ρΓm(a), Γm(a, 0) = Γm(a),
Γm(a,−ρ) =
(−1)r Γm(a)(−a +m + 1/2)ρ
, Re(a) > r1 +m − 12
,(4.3)
where the generalized hypergeometric coefficient (a)ρ is defined by
(a)ρ =m∏i=1
(a − i − 1
2
)
ri
. (4.4)
Further, det(Im −X)−a, in terms of zonal polynomials, can be expanded as
det (Im −X)−a =∞∑k=0
∑κ
(a)κCκ(X)k!
, ‖X‖ < 1, (4.5)
where∑
κ denotes summation over all ordered partitions κ of k.
14 ISRN Probability and Statistics
For properties and further results, the reader is referred to Constantine [7] and Guptaand Nagar [2].
Lemma 4.1. Let T be an m ×m arbitrary complex symmetric matrix. Then
∫ Im
0det (R)a−(m+1)/2 det (Im + R)−(a+b)Cκ(RT)dR
=Γm(a, κ)Γm(b,−κ)
Γm(a + b)Cκ(T)
=(−1)kΓm(a, κ)Γm(b)
(−b + (m + 1)/2)κΓm(a + b)Cκ(T), Re(b) > k1 +
m − 12
.
(4.6)
Davis [9, 10] introduced a class of polynomials Cκ,λφ
(X,Y ) of m × m symmetricmatrix arguments X and Y , which are invariants under the transformation (X,Y ) →(HXHT,HYHT ), H ∈ O(m). For properties and applications of invariant polynomials, werefer to Davis [9, 10], Chikuse [11], and Nagar and Gupta [12]. Let κ, λ, φ, and ρ be orderedpartitions of the nonnegative integers k, , f = k + , and r, respectively. Then
Cκ,λφ (X,X) = θκ,λ
φCφ(X), θκ,λ
φ=
Cκ,λφ (Im, Im)
Cφ(Im),
Cκ,λφ (X, Im) = θκ,λ
φ
Cφ(Im)Cκ(X)Cκ(Im)
,
(4.7)
Cκ,0κ (X,Y ) ≡ Cκ(X), C0,λ
λ (X,Y ) ≡ Cλ(Y ),
Cκ(X)Cλ(Y ) =∑φ∈κ·λ
θκ,λφ Cκ,λ
φ (X,Y ),(4.8)
where φ ∈ κ · λ denotes that irreducible representation of Gl(m,R), the group of m × m realinvertible matrices, indexed by 2φ, appears in the decomposition of the tensor product 2κ⊗2λof the irreducible representation indexed by 2κ and 2λ. Further,
∫ Im
0det (R)t−(m+1)/2 det (Im − R)u−(m+1)/2Cκ,λ
φ (AR,A(Im − R))dR
=Γm(t, κ)Γm(u, λ)Γm
(t + u, φ
) θκ,λφ
Cφ(A),
(4.9)
∫ Im
0det (R)t−(m+1)/2 det (Im − R)u−(m+1)/2Cκ,λ
φ (AR,BR)dR
=Γm
(t, φ
)Γm(u)
Γm(t + u, φ
) Cκ,λφ (A,B).
(4.10)
ISRN Probability and Statistics 15
From the density of V , we have
E[Cκ(BV )] =Γm
[β + (m + 1)/2
]
Γm(β)Γm[m + 1/2]
∫
V>0Cκ(VB)det (Im + V )−β−(m+1)/2 dV
=(−1)k((m + 1)/2)κ(−β + (m + 1)/2
)κ
Cκ(B), Re(β)> k1 +
m − 12
,
(4.11)
where the last line has been obtained by using (4.6).Using results (4.1)–(4.2) on zonal polynomials, it is easy to see that
E[C(1)(BV )
]=
m + 12β −m − 1
C(1)(B),
E[tr(BV )] =m + 1
2β −m − 1tr(B), β >
m + 12
,
E[C(2)(BV )
]=
(m + 1)(m + 3)(2β −m − 1
)(2β −m − 3
)C(2)(B)
=(m + 1)(m + 3)
3(2β −m − 1
)(2β −m − 3
)[(trB)2 + 2 tr
(B2
)], β >
m + 32
,
E[C(12)(BV )
]=
m(m + 1)(2β −m − 1
)(2β −m
)C(12)(B)
=2m(m + 1)
3(2β −m − 1
)(2β −m
)[(trB)2 − tr
(B2
)], β >
m + 12
,
E[C(3)(BV )
]=
(m + 1)(m + 3)(m + 5)(2β −m − 1
)(2β −m − 3
)(2β −m − 5
)C(3)(B)
=(m + 1)(m + 3)(m + 5)
15(2β −m − 1
)(2β −m − 3
)(2β −m − 5
)
×[(trB)3 + 6(trB)
(trB2
)+ 8 tr
(B3
)], β >
m + 52
,
E[C(2,1)(BV )
]=
m(m + 1)(m + 3)(2β −m
)(2β −m − 1
)(2β −m − 3
)C(2,1)(B)
=3m(m + 1)(m + 3)
15(2β −m
)(2β −m − 1
)(2β −m − 3
)
×[(trB)3 + (trB)
(trB2
)− 2 tr
(B3
)], β >
m + 32
,
16 ISRN Probability and Statistics
E[C(13)(BV )
]=
(m − 1)m(m + 1)(2β −m
)(2β −m + 1
)(2β −m − 1
)C(13)(B)
=(m − 1)m(m + 1)
3(2β −m + 1
)(2β −m
)(2β −m − 1
)
×[(trB)3 − 3(trB)
(trB2
)+ 2 tr
(B3
)], β >
m + 12
,
E[tr (BV )2
]= E
[C(2)(BV )
] − 12E[C(12)(BV )
]
=2β(m + 1)(
2β −m)(2β −m − 1
)(2β −m − 3
) (trB)2
+(m + 1)
[(m + 2)
(2β −m − 1
)+ 2
](2β −m
)(2β −m − 1
)(2β −m − 3
)(trB2
), β >
m + 32
,
(4.12)
E[tr (BV )3
]= E
[C(3)(BV )
] − 14E[C(2,1)(BV )
]+14E[C(13)(BV )
]
=1 +m(
2β −m + 1)(2β −m
)(2β −m − 1
)(2β −m − 3
)(2β −m − 5
)
×[110
{(m − 1)m(3 +m)(5 +m) − 4
(m3 + 5m2 +m − 5
)β
+4(m2 + 3m + 10
)β2}(trB)3
+110
{(m − 1)m(3 +m)(5 +m)
−4(m(3 +m)(17 +m) − 30)β + 4(m2 + 33m + 60
)β2}(trB)
(trB2
)
+45
{(m − 1)m(3 +m)(5 +m) + 20β −m
(4m2 + 25m + 29
)β
+2(2m2 + 11m + 20
)β2}tr(B3
)], β >
m + 52
,
E[tr(BV ) tr (BV )2
]= E
[C(3)(BV )
]+16E[C(2,1)(BV )
] − 12E[C(13)(BV )
]
=1 +m(
2β −m + 1)(2β −m
)(2β −m − 1
)(2β −m − 3
)(2β −m − 5
)
×[− 115
{(m − 1)m(3 +m)(5 +m) − 2
(2m3 − 5m2 − 48m + 15
)β
+4(m2 − 12m − 15
)β2}(trB)3
ISRN Probability and Statistics 17
+215
{7(m − 1)m(3 +m)(5 +m)
+90β − 2m(14m2 + 70m + 39
)β + 4
(7m2 + 21m + 45
)β2}
× (trB)(trB2
)
+215
{(m − 1)m(3 +m)(5 +m) − 4(m(3 +m)(17 +m) − 30)β
+4(m2 + 33m + 60
)β2}tr(B3
)], β >
m + 52
.
(4.13)
Further, using the invariance of the distribution of V and the above results, one obtains
E(V ) =m + 1
2β −m − 1Im, β >
m + 12
,
E(V 2
)=
(m + 1)[4(m + 1)β −m(m + 3)
](2β −m
)(2β −m − 1
)(2β −m − 3
)Im, β >m + 32
,
E(V 3
)=
1 +m
10(2β −m + 1
)(2β −m
)(2β −m − 1
)(2β −m − 3
)(2β −m − 5
)
×[(m − 1)m(3 +m)(5 +m)
(m2 +m + 8
)
− 4(m5 + 6m4 + 29m3 + 96m2 + 28m − 40
)β
+4(m4 + 4m3 + 51m2 + 104m + 80
)β2]Im, β >
m + 52
,
E[(
trV 2)V]= − 1 +m
15(2β −m + 1
)(2β −m
)(2β −m − 1
)(2β −m − 3
)(2β −m − 5
)
×[(m − 1)m(3 +m)(5 +m)
(m2 − 14m − 2
)
+ 2(−2m5 + 33m4 + 192m3 + 143m2 + 114m − 120
)β
+4(m4 − 26m3 − 59m2 − 156m − 120
)β2]Im, β >
m + 52
.
(4.14)
18 ISRN Probability and Statistics
Theorem 4.2. Let V1 and V2 be independent, Vi ∼ Pm(βi), i = 1, 2. Define S = V1 + V2 and R =(V1 + V2)
−1/2V2(V1 + V2)−1/2. Then, the density of S is given by
Γm[β1 + (m + 1)/2
]Γm
[β2 + (m + 1)/2
]
Γm(β1)Γ2m[(m + 1)/2]Γm
(β2) det (S)(m+1)/2 det (Im + S)−(β1+β2+m+1)
×∞∑k=0
∞∑l=0
∑κ
∑λ
(β1 + (m + 1)/2
)κ
(β2 + (m + 1)/2
)λ
k! l!
×∑φ∈κ·λ
(θκ,λφ
)2 Γ((m + 1)/2, κ)Γ((m + 1)/2, λ)Γ(m + 1, φ
) Cφ
((Im + S)−1S
), S > 0.
(4.15)
Further, the density of R is derived as
Γm[β1 + (m + 1)/2
]Γm
[β2 + (m + 1)/2
]
Γm(β1)Γm
(β2)Γ2m[(m + 1)/2]
×∞∑k=0
∞∑l=0
∑κ
∑λ
(β1 + (m + 1)/2
)κ
(β2 + (m + 1)/2
)λ
k! l!
×∑φ∈κ·λ
Γ(m + 1, φ
)Γm
(β1 + β2
)
Γm(β1 + β2 +m + 1, φ
) θκ,λφ
Cκ,λφ (R, Im − R), 0 < R < Im.
(4.16)
Proof. The joint density of V1 and V2 is given by
Γm[β1 + (m + 1)/2
]Γm
[β2 + (m + 1)/2
]
Γm(β1)Γm
(β2)Γ2m[(m + 1)/2]
× det (Im + V1)−β1−(m+1)/2 det (Im + V2)−β2−(m+1)/2, V1 > 0, V2 > 0.
(4.17)
Making the transformation V2 = S1/2RS1/2 and V1 = S1/2(Im − R)S1/2 with the JacobianJ(V1, V2 → S,R) = det(S)(m+1)/2 in (4.17), the joint density of R and S is derived as
Γm[β1 + (m + 1)/2
]Γm
[β2 + (m + 1)/2
]
Γm(β1)Γm
(β2)Γ2m[(m + 1)/2]
det (S)(m+1)/2 det (Im + S)−(β1+β2+m+1)
× det (Im − S1R)−β1−(m+1)/2 det (Im − S1(Im − R))−β2−(m+1)/2,
(4.18)
ISRN Probability and Statistics 19
where S1 = (Im + S)−1S, 0 < R < Im, and S > 0. Since, ‖S1R‖ < 1 and ‖S1(Im − R)‖ < 1, using(4.5), we can write
det (Im − S1R)−β1−(m+1)/2 =∞∑k=0
∑κ
(β1 + (m + 1)/2
)κ
k!Cκ(S1R),
det (Im − S1(Im − R))−β2−(m+1)/2 =∞∑l=0
∑λ
(β2 + (m + 1)/2
)λ
l!Cλ(S1(Im − R)),
(4.19)
where κ and λ are the ordered partitions of k and l, respectively. Now, the application of (4.8)yields
det (Im − S1R)−β1−(m+1)/2 det (Im − S1(Im − R))−β2−(m+1)/2
=∞∑k=0
∞∑l=0
∑κ
∑λ
(β1 + (m + 1)/2
)κ
(β2 + (m + 1)/2
)λ
k! l!
×∑φ∈κ·λ
θκ,λφ
Cκ,λφ (S1R, S1(Im − R)).
(4.20)
Finally, substituting (4.20) in (4.18), the joint density of R and S is obtained as
Γm[β1 + (m + 1)/2
]Γm
[β2 + (m + 1)/2
]
Γm(β1)Γm
(β2)Γ2m[(m + 1)/2]
det (S)(m+1)/2 det (Im + S)−(β1+β2+m+1)
×∞∑k=0
∞∑l=0
∑κ
∑λ
(β1 + (m + 1)/2
)κ
(β2 + (m + 1)/2
)λ
k! l!
×∑φ∈κ·λ
θκ,λφ
Cκ,λφ (S1R, S1(Im − R)), 0 < R < Im, S > 0.
(4.21)
Now, the integration of R in (4.21) using (4.9) yields the density of S. The density of R isobtained by substituting S1 = (Im + S)−1S with the Jacobian J(S → S1) = det(I − S1)
−(m+1) in(4.21) and integrating S1 by using (4.10).
Acknowledgment
The research work of D. K. Nagar was supported by the Comite para el Desarrollo de laInvestigacion, Universidad de Antioquia, research Grant no. IN560CE.
References
[1] N. L. Johnson, S. Kotz, and N. Balakrishnan, Continuous Univariate Distributions, John Wiley & Sons,New York, NY, USA, 2nd edition, 1994.
[2] A. K. Gupta and D. K. Nagar,Matrix Variate Distributions, vol. 104, Chapman & Hall, Boca Raton, Fla,USA, 2000.
20 ISRN Probability and Statistics
[3] L. Cardeno, D. K. Nagar, and L. E. Sanchez, “Beta type 3 distribution and its multivariategeneralization,” Tamsui Oxford Journal of Mathematical Sciences, vol. 21, no. 2, pp. 225–241, 2005.
[4] A. K. Gupta and D. K. Nagar, “Matrix-variate beta distribution,” International Journal of Mathematicsand Mathematical Sciences, vol. 24, no. 7, pp. 449–459, 2000.
[5] A. K. Gupta and D. K. Nagar, “Properties of matrix variate beta type 3 distribution,” InternationalJournal of Mathematics and Mathematical Sciences, vol. 2009, Article ID 308518, 18 pages, 2009.
[6] C. S. Herz, “Bessel functions of matrix argument,” Annals of Mathematics, vol. 61, pp. 474–523, 1955.[7] A. G. Constantine, “Some non-central distribution problems in multivariate analysis,” Annals of
Mathematical Statistics, vol. 34, pp. 1270–1285, 1963.[8] A. T. James, “Distributions of matrix variates and latent roots derived from normal samples,” Annals
of Mathematical Statistics, vol. 35, pp. 475–501, 1964.[9] A. W. Davis, “Invariant polynomials with two matrix arguments extending the zonal polynomials:
applications to multivariate distribution theory,” Annals of the Institute of Statistical Mathematics, vol.31, no. 3, pp. 465–485, 1979.
[10] A. W. Davis, “Invariant polynomials with two matrix arguments, extending the zonal polynomials,”in Multivariate Analysis-V, P. R. Krishnaiah, Ed., pp. 287–299, North-Holland, Amsterdam, TheNetherlands, 1980.
[11] Y. Chikuse, “Distributions of some matrix variates and latent roots in multivariate Behrens-Fisherdiscriminant analysis,” The Annals of Statistics, vol. 9, no. 2, pp. 401–407, 1981.
[12] D. K. Nagar and A. K. Gupta, “Matrix-variate Kummer-beta distribution,” Journal of the AustralianMathematical Society, vol. 73, no. 1, pp. 11–25, 2002.